Zaman serisi tahmini için otomatik çok modelli yaklaşım

In this thesis, we present a multi-model approach for time series forecasting. Our framework is automated, easy to use and spans around short-medium terms forecast horizons. We focus on multiple basic and advance time series forecasting models. Our basic models include simple moving average, simple exponential smoothing, Holt's and winter's model. We optimize the parameters used in these models, and for advance models we extend Box-Jenkins methodology to automated Auto- Regressive Integrated Moving Average (ARIMA) and Seasonal Auto-Regressive Integrated Moving Average (SARIMA) models. The ARIMA and SARIMA models are complex models and need expert judgement and iterative procedures to select a best fitting model.We substitute expert judgement with statistical tests and iterative procedures with automated enumeration technique. The best fitting model is selected based on the results of a number of statistical and error estimation tests. We tested our system on M-competition data set provided by International Institute of Forecasters. The dataset is comprised of multiple time series from social and economic backgrounds. From our experimental work, we conclude that our SARIMA model outperforms all other models with an average MAPE of <1% for 1-period ahead and approximately 9% for 6 periods-ahead forecast horizons.